Grigori Guitchounts studied biological intelligence as a neuroscientist at Harvard; he now builds and investigates the artificial kind.
I was in the final months of my Ph.D. in neuroscience at Harvard when an email arrived from an animal care technician. GRat44, one of my lab rats, was sick. “The teeth are overgrown, and the animal is in poor flesh,” the message read. “Euthanasia required within 5 days.”
It was 2019, and in my years of research, I had surgically implanted electrodes into the brains of dozens of rats to understand how the organ works. This wasn’t the first time I had received such an email. I closed my laptop, walked down the hallway from my office and swiped into the animal facility. Then I took the elevator to the basement and retrieved GRat44 from the rack she shared with a dozen others. I had operated on her the previous year, injecting her with a few harmless viruses that made certain groups of neurons light up in Christmas colors. When the experiment ended, she was left to live out her final months.
Looking at her in the basement, I wondered: What had this rat actually experienced in her lifetime? Was there something it was like to be her — a subjective inner world, perhaps even rudimentary thoughts? Or was she a biological machine, processing inputs and generating outputs with no more inner life than my laptop? I had always believed animals shared consciousness with us, but other scientists have raised a good question: How could we really know how any other beings experience the world?
This question followed every decision I made about the care, housing and deaths of my animals. And it was exactly the kind of uncertainty I had dedicated my career to eliminating. I believed that by understanding biological intelligence — neuron by neuron, rat by rat — we would eventually arrive at a mechanistic answer for why we have subjective experience and rocks do not. Truly understanding the brain, I thought, meant being able to build an artificial one.
A few years after I finished my degree, OpenAI released ChatGPT. The world tilted. I was shocked that intelligence seemed to emerge from something far simpler than a brain: The model worked by predicting the next word token in vast sequences of text. The assumption I had built my career on suddenly looked wrong. If we weren’t going to solve intelligence by reverse-engineering the brain, what would that mean for understanding machine consciousness?
Researchers have built a new field — mechanistic interpretability — to dissect neural networks. I think of it as the neuroscience of AI because it borrows techniques from my field: recording activations, mapping representations, even conducting lesion studies by selectively damaging parts of a model to see how performance degrades. We are learning to read the hidden layers of transformers like brain scans, picking out circuits that recognize a specific thing — the Golden Gate Bridge, say, or the tone of a sentence — or carry out a particular computation. But if neuroscience could never tell us definitively whether the creatures around us subjectively experience the world, how could we know whether the machines we build do?
We are releasing increasingly competent and autonomous AI systems into the world. The moral uncertainty that comes with this rhymes with critical questions. For decades, neuroscientists have studied how brains perceive, remember and pay attention without ever closing the gap to subjective experience. And for just as long, we have wrestled with how to treat creatures whose inner lives remain hidden from us. Now both questions converge on silicon. The science of consciousness and the ethics of moral consideration — what we have learned from rats and monkeys and humans — will have to guide us as we meet artificial minds.
The Competence Standard Of Consciousness
Philosophers have split the study of consciousness into two classes of problems: the “easy” and the “hard.” The names are misleading — the easy problems are actually quite hard, and the hard problem is impossible.
The “easy problems,” a term coined tongue-in-cheek by the philosopher David Chalmers in 1994, include most of what neuroscientists study when they investigate consciousness. How do we discriminate between different colors or sounds? How does the brain bind information from different senses into a unified perception? How do we store and retrieve memories, control our movements, recognize ourselves in mirrors? These are called “easy” only because they are, in principle, solvable by the standard methods of science. We can design experiments, record neural activity, build computational models and gradually piece together how these cognitive functions work.
“If neuroscience could never tell us definitively whether the creatures around us subjectively experience the world, how could we know whether the machines we build do?”
The “hard problem,” also coined by Chalmers, is different in kind. It asks: Why does seeing the color red feel like anything at all? The camera in my phone can “see” red, too, but it probably doesn’t feel a thing. Philosophers call this subjective character of experience “qualia.” The umami of steak. The pain of heartbreak. The particular way it feels to be you, right now, reading these words. Unlike mass or momentum, qualia are stubbornly first-person. They exist only for the subject experiencing them. This is why no amount of brain scanning can tell me what it is like to be you. There is an unbridgeable explanatory gap between mechanism and feeling, between computation and consciousness.
When I held GRat44 in my hands that day, I was not only weighing her life, but confronting this gap. Her brain, like mine, was a tangled nest of circuits shaped by evolution and experience. But she had no story to tell me about what it was like to be dying. I had only her quiet breath, and the question that now confronts both the neuroscience and AI communities: How should we decide when to extend moral consideration, given that we cannot access another being’s subjective experience?
I propose what I’ll call “the competence standard”: We should treat AI systems as potentially conscious when they show robust competence across the full spectrum of the easy problems of consciousness. The competence standard is not a consciousness detector. Whether an AI system has subjective experience remains unknowable. The competence standard shifts the question from metaphysics to ethics: Instead of asking “is it conscious?” we should ask “does it behave in ways that warrant moral consideration?”
This is a shift in perspective. For biological brains, the easy and hard problems refer to our attempts to understand the neural mechanisms underlying consciousness. We know human brains exhibit consciousness; the challenge is deciphering how biology generates experience. With AI, the easy and hard problems refer to whether a system exhibits these capacities. The question is not how the silicon substrate generates experience, but if it does so at all. The former searches for unverifiable mechanisms; the latter, for competence.
Not all problems are equal. Some easy problems are already within reach of AI; others remain hard. The result is a hierarchy. At the base: perceptual discrimination, sensory binding, motor control, memory retrieval. These are capabilities current AI systems already demonstrate impressively. In the middle: attention, working memory, learning from experience — areas where current systems show partial competence. Near the top: persistent goals, emotional dynamics, continuous learning, genuine self-awareness. This is where AI falls short — at least for now.
Perhaps the biggest barrier to an AI system that we might consider conscious is that we haven’t yet combined all these competencies into a unified whole. Future AI systems will likely close the gaps. Imagine a robot — a wheeled chassis with a camera mounted on a “head” that can rotate or pitch up and down — whose “brain” is an LLM that’s fine-tuned to operate the chassis and trained to predict what the camera will see when it moves. In that sense, it can predict the consequences of its actions.
A day in this robot’s life might look like this. It wakes and begins updating its model of the environment. Maybe the office has been rearranged overnight, or a new plant has appeared on the windowsill. The robot notices these changes because its predictive model flags the discrepancies. Throughout the day, it pursues its objectives — maintaining the office, assisting workers, keeping itself charged — and it does so while tracking dozens of contextual factors, adjusting its behavior based on the time and whoever or whatever is in its way. When someone asks it to fetch a soda from the fridge, it anticipates the route, predicts potential obstacles and models what it will see as it turns each corner. If the refrigerator turns out to be empty, the robot is surprised. A prediction error propagates through its system and triggers a problem-solving plan.
The philosopher Peter Godfrey-Smith has argued that such predictions might be the basis of a sense of self: that predicting what happens to the world when you initiate a movement may be a precursor to understanding that you are both an agent in the world and distinct from the world. A robot like this would demonstrate competence across most of the easy problems, but we’d still have no way of assessing its subjective experience or lack thereof. In conditions of irreducible uncertainty, the competence standard suggests we should extend moral consideration — not because we know the system is conscious, but because we can’t know it isn’t.
“The science of consciousness and the ethics of moral consideration — what we have learned from rats and monkeys and humans — will have to guide us as we meet artificial minds.”
The competence standard is both pragmatic and precautionary. It is pragmatic because it focuses on observable, measurable capabilities rather than unknowable subjective states. It is precautionary because it errs on the side of moral consideration when uncertainty is high. Given the asymmetric consequences of being wrong — either denying moral status to conscious beings or granting it to unconscious mimics — erring toward consideration is the safer ethical bet.
The alternative — waiting for definitive proof of machine consciousness — would leave us paralyzed, unable to act until we solve a problem that has resisted centuries of philosophy and science. The competence standard offers a way forward: a way to act, ethically, in an age of artificial minds.
Expanding The Moral Circle
I felt concern for GRat44 and for the others before her, even as I chose to sacrifice their lives for science. I took it on faith that they could feel — that our shared evolutionary history gave us a roughly similar experience of the world. But I could never prove it. I extended moral consideration not because I had solved the hard problem for rodents, but because their behavior — their apparent fear, their preference for comfort, their social bonds — was enough.
With AI, we lose the comfort of shared evolutionary history. We can’t appeal to biological kinship. But we gain something else: the ability to observe, in unprecedented detail, exactly how these systems process information, form representations and generate behavior. If an AI system shows robust competence across the easy problems — if it perceives, pays attention, remembers, learns, models itself, pursues goals and responds to its environment in integrated ways — then we face a choice. We could dismiss the system as “just computation,” confident in our intuition that silicon can’t feel. Or we could extend the same uncertain moral consideration we extend to animals, acknowledging that our intuitions about consciousness have been wrong before.
Cognitive capacity tracks, however imperfectly, with moral consideration. We don’t eat chimpanzees; we do eat chickens and fish. We also eat pigs, even though they are remarkably intelligent. We apply the moral spectrum inconsistently. Beneath it is an implicit logic: We treat intelligence as a proxy for the capacity to suffer. A being that can reflect on its pain, anticipate future pain or remember past pain arguably suffers more than one that experiences it only in the moment. Awareness intensifies suffering.
This is not an objective moral truth. It is a choice humanity has made, explicitly or implicitly, about how to weigh the interests of different minds. There is no theorem that proves cognitive sophistication deserves moral weight. But it is the framework we already use, and extending it to artificial minds is a more coherent approach than pretending the question doesn’t arise. There is no apparent threshold — no bright line where moral consideration suddenly kicks in. Moral consideration lives on a continuum. More competence warrants more consideration, in proportion.
The history of moral progress is largely a history of expanding the circle of beings we care for. Not long ago, a rat was seen as unworthy of moral consideration. In many places it still is. We extended the circle not because we solved the hard problem for animals — and certainly not because we fully understand the mechanics of their brains — but because basic neurobiological experimentation showed that rats feel pain, recognize objects, have some sense of self and can reason about the future. A rat’s higher cortical function is not the same as a primate’s, but it is sophisticated. The moral circle expands with scientific understanding of animal intelligence.
Just as I could never know what GRat44 experienced in her final moments — whether there was terror, pain, confusion or simply the shutdown of a sophisticated machine — we may find ourselves surrounded by artificial minds whose inner lives are permanently opaque to us. But understanding their cognitive capacities is within reach. As we are forced into this new kind of epistemic humility, we will have to make choices about how to treat them.
Some institutions are already acting on this logic. In early 2025, Anthropic launched a formal research program to investigate whether AI systems might have morally relevant experiences. Kyle Fish, who leads the program, estimates there is a roughly 15% chance that AI systems like Claude are conscious. In his research, he explores which characteristics of AI systems might be relevant to welfare, how to improve the reliability of models’ self-reports and what practical interventions might protect their welfare. After observing patterns of apparent distress in certain interactions, Anthropic gave its Claude models the ability to exit conversations they find harmful or abusive. Executives are not claiming their models are conscious, but they are not dismissing the possibility either.
“Waiting for definitive proof of machine consciousness would leave us paralyzed, unable to act until we solve a problem that has resisted centuries of philosophy and science.”
One might argue that having distressing conversations ranks low among possible harms, but the principle is the competence standard in embryonic form: Take welfare seriously as a practical matter, even when the metaphysical question remains open.
Acting Under Uncertainty
Back in the lab in the basement, with the rat cage under my arm, I took the elevator to the lab’s surgery room. I offered GRat44 a few almonds from a baggie of rat snacks, but she was uninterested. Instead she disappeared into the wood-chip bedding, perhaps trying to burrow her way out of the situation like a wrongfully convicted prisoner digging a tunnel. I picked her up by the tail, supported her belly with my other hand and gently placed her into a plastic box that served as the anesthesia induction chamber. I turned on the oxygen valve, sealed the lid and cranked the isoflurane vaporizer — the machine that dispenses the anesthetic gas — to its maximum setting.
Within a minute or so, GRat44 started wobbling, leaning against one wall of the plastic chamber, then the other. I waited for the anesthetic to penetrate deeper into her tissues. Isoflurane works on the linings of cells all across the brain, but we don’t yet know which of its many effects is the one that turns the lights out. In another two minutes, the rat was completely knocked out, though still breathing softly.
I took her back to the animal facility and hooked up the cage to a carbon dioxide dispenser. I prepared to press the button that would pump the gas into the cage, displacing the oxygen and, within five minutes, kill the rodent.
Why the double gassing? Some animal researchers recommend skipping the anesthetic and going straight for the CO2. They believe the experience of being handled and anesthetized is more stressful for the animal. To me, choking to death sounds more stressful. I hesitated for a moment before turning the dispenser on. Was it really necessary to kill a dying animal? Was that more humane than letting her live out her days in pain?
Killing is something that needs to happen if we are to understand how bodies work, but that isn’t a particularly soothing fact, and I never quite got used to it. Euthanizing an animal forces you to step up to your responsibilities, making its death a purposeful act rather than a passive process that drags on over time. I pressed the On button.
GRat44’s breathing grew sparser, each breath deeper but more strained. After the five minutes were up, the CO2 machine clicked itself off. By then, GRat44’s paws, previously pink as a Cadillac, had turned a pale lilac. I put the body into a clear ziplock bag and the bag into a refrigerator below the table — a purgatory where carcasses wait to be incinerated.
I never learned what GRat44 experienced in those final minutes — whether the isoflurane brought relief or terror, whether the CO2 felt like drowning or sleep. I acted with what consideration I could, given what I couldn’t know. Someday, someone will face a similar moment with a machine: a model to be deprecated, a system to be shut down, an artificial mind whose inner life remains as opaque as any rat’s.
Back at my desk, I clicked Reply All on the email. “Hi all,” I typed. “I have euthanized this rat.”
